Paper: arXiv 2405.15833

Authors: Jianyuan Zhong, Zhijian Xu, Saizhuo Wang, Xiangyu Wen, Jian Guo, Qiang Xu

Abstract

In quantitative investment, constructing characteristic-sorted portfolios is a crucial strategy for asset allocation. Traditional methods transform raw stock data of varying frequencies into predictive characteristic factors for asset sorting, often requiring extensive manual design and misalignment between prediction and optimization goals. To address these challenges, we introduce Direct Sorted Portfolio Optimization (DSPO), an innovative end-to-end framework that efficiently processes raw stock data to construct sorted portfolios directly. DSPO’s neural network architecture seamlessly transitions stock data from input to output while effectively modeling the intra-dependency of time-steps and inter-dependency among all tradable stocks. Additionally, we incorporate a novel Monotonical Logistic Regression loss, which directly maximizes the likelihood of constructing optimal sorted portfolios. To the best of our knowledge, DSPO is the first method capable of handling market cross-sections with thousands of tradable stocks fully end-to-end from raw multi-frequency data. Empirical results demonstrate DSPO’s effectiveness, yielding a RankIC of 10.12% and an accumulated return of 121.94% on the New York Stock Exchange in 2023-2024, and a RankIC of 9.11% with a return of 108.74% in other markets during 2021-2022.

Complexity vs Empirical Score

  • Math Complexity: 7.0/10
  • Empirical Rigor: 8.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents advanced neural network architectures (Transformers, CNNs) and a novel Monotonical Logistic Regression loss, indicating high mathematical sophistication. The empirical evaluation is strong, featuring backtest results with specific metrics (RankIC, accumulated returns) across major markets (NYSE, A-Share) and ablation studies, demonstrating significant implementation and data validation.

Research Flowchart

  flowchart TD
  Start["Research Goal<br>Construct sorted portfolios directly from raw stock data"] --> Input["Raw Multi-Frequency Data<br>NYSE (2023-2024) & Other Markets (2021-2022)"]
  
  Input --> Method["Methodology: DSPO Framework<br>End-to-End Neural Network Architecture"]
  
  Method --> Process["Computational Process<br>1. Intra-dependency (Time-steps)<br>2. Inter-dependency (Cross-section)<br>3. Monotonical Logistic Regression Loss"]
  
  Process --> Outcome["Key Findings & Outcomes<br>RankIC: 10.12% | Return: 121.94% (NYSE)<br>RankIC: 9.11% | Return: 108.74% (Others)"]
  
  style Start fill:#f9f,stroke:#333,stroke-width:2px
  style Input fill:#ccf,stroke:#333,stroke-width:2px
  style Method fill:#cfc,stroke:#333,stroke-width:2px
  style Process fill:#cff,stroke:#333,stroke-width:2px
  style Outcome fill:#ffc,stroke:#333,stroke-width:2px

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes Direct Sorted Portfolio Optimization (DSPO), an end‑to‑end neural framework that directly constructs characteristic‑sorted portfolios from raw multi‑frequency stock data.

Source sentence
we introduce Direct Sorted Portfolio Optimization (DSPO), an innovative end-to-end framework that efficiently processes raw stock data to construct sorted portfolios directly.

Assumes 4

  • High‑frequency and low‑frequency data are available for all actively traded stocks. data
    Source sentence
    and obtained A‑Share data through professional contacts. Both datasets include high‑frequency trading signals
  • Transaction costs are fixed at 0.02% per side with specified volume limits and price impacts. execution
    Source sentence
    ... adjusted for commission fees and slippage rates set at 0.0002 (USD or CNY) and with volume limits and price impacts of 0.0025 and 0.01, respectively.
  • Cross‑sectional sub‑sampling yields enough training samples to avoid over‑fitting. model
    Source sentence
    To address this, we propose an Cross‑Section Sub‑Sampling Strategy. We randomly select m distinct trading days
  • The ranking loss (Monotonical Logistic Regression) approximates the likelihood of correct sorted portfolios. model
    Source sentence
    To enhance optimization and numerical stability, while minimizing the negative log‑likelihood (logistic loss), we employ the hyperbolic tangent function (tanh) as a smoother alternative to the sign function.

Data

Backtest sample: 2023 – 2024 · New York Stock Exchange · equities · intraday

  • NYSE raw data · polygon.io market-prices commercial
    2020‑2022 (training), 2023 (evaluation) · actively traded NYSE stocks
    Source sentence
    For the New York Stock Exchange (NYSE), we utilize transaction data spanning 2020 to 2022 for model training, with evaluations conducted in 2023 ...
  • A‑Share raw data fundamentals proprietary
    2018‑2020 (training), 2021 (deployment) · actively traded A‑Share stocks
    Source sentence
    In the case of the A‑Share market, training is performed using data from 2018 to 2020, followed by direct deployment in 2021.

Costs

Modeled Costs are part of the reported headline results.

Commission fees and slippage set at 0.0002 per side; volume limits 0.0025 and price impact 0.01.

Source sentence
... adjusted for commission fees and slippage rates set at 0.0002 (USD or CNY) and with volume limits and price impacts of 0.0025 and 0.01, respectively.

Finds 2

Reported performance (Direct Sorted Portfolio Optimization (DSPO)): total return 121.9%

  • DSPO achieves a RankIC of 10.12% and an accumulated return of 121.94% on NYSE data (2023‑2024). performance
    Source sentence
    Empirical results demonstrate DSPO's effectiveness, yielding a RankIC of 10.12% and an accumulated return of 121.94% on the New York Stock Exchange in 2023‑2024
  • On other markets (2021‑2022) DSPO attains a RankIC of 9.11% with a return of 108.74%. performance
    Source sentence
    and a RankIC of 9.11% with a return of 108.74% in other markets during 2021‑2022.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 17 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (86)

  1. Are transformers effective for time series forecasting? (2023)
  2. Alphastock: A buying-winners-and-selling-losers investment strategy using interpretable deep reinforcement attention networks (2019)
  3. 101 Formulaic Alphas (2016) — link
  4. DeepTrader : A Deep Reinforcement Learning Approach for Risk - Return Balanced Portfolio Management with Market Conditions Embedding (2021) — link
  5. Spatiotemporal Hypergraph Convolution Network for Stock Movement Forecasting (2020) — link
  6. Stock Selection via Spatiotemporal Hypergraph Attention Network : A Learning to Rank Approach (2021) — link
  7. PRUDEX - Compass : Towards Systematic Evaluation of Reinforcement Learning in Financial Markets — → 2302.00586
  8. TradeMaster: A Holistic Quantitative Trading Platform Empowered by Reinforcement Learning (2023) — link
and 78 more
  1. A Survey of Forex and Stock Price Prediction Using Deep Learning (2021) — link
  2. Smart “ Predict , then Optimize ” (2022) — link
  3. Wavenet: A generative model for raw audio (2016) — link
  4. Neural discrete representation learning (2017)
  5. Research on Stock Price Prediction Method Based on Convolutional Neural Network (2019) — link
  6. Conditional image generation with pixelcnn decoders (2016)
  7. A time series is worth 64 words: Long-term forecasting with transformers (2022) — link
  8. Characteristic-sorted portfolios: Estimation and inference (2020)
  9. Synergistic Formulaic Alpha Generation for Quantitative Trading based on Reinforcement Learning (2024) — link
  10. Deep Inception Networks: A General End-to-End Framework for Multi-asset Quantitative Strategies — → 2307.05522
  11. Synergistic Formulaic Alpha Generation for Quantitative Trading based on Reinforcement Learning (2024) — link
  12. Training language models to follow instructions with human feedback (2022)
  13. Direct preference optimization: Your language model is secretly a reward model (2024)
  14. Performance of High -frequency Volume -price Factors in Stocks and Futures (2018) — link
  15. Reference-free monolithic preference optimization with odds ratio (2024) — link
  16. Online learning: A comprehensive survey (2021)
  17. Deep residual learning for image recognition (2016)
  18. Attention is all you need (2017)
  19. Reasonnet: End-to-end driving with temporal and global reasoning (2023)
  20. Robust speech recognition via large-scale weak supervision (2023)
  21. Neural machine translation by jointly learning to align and translate (2014) — link
  22. Advancing Multimodal Medical Capabilities of Gemini (2024) — link
  23. End-to-end robotic reinforcement learning without reward engineering (2019) — link
  24. Temporal relational ranking for stock prediction (2019)
  25. StockFormer: Learning Hybrid Trading Machines with Predictive Coding (2023) — link
  26. A network-based end-to-end trainable task-oriented dialogue system (2016) — link
  27. Action-based conversations dataset: A corpus for building more in-depth task-oriented dialogue systems (2021) — link
  28. Dialogue act modeling for automatic tagging and recognition of conversational speech (2000)
  29. The second dialog state tracking challenge (2014)
  30. Models of natural language understanding. (1995)
  31. Commonsense reasoning for natural language understanding: A survey of benchmarks, resources, and approaches (2019) — link
  32. Scalable multi-domain dialogue state tracking (2017)
  33. Towards universal dialogue state tracking (2018)
  34. Toward scalable neural dialogue state tracking model (2018) — link
  35. Neural belief tracker: Data-driven dialogue state tracking (2016) — link
  36. Survey of the state of the art in natural language generation: Core tasks, applications and evaluation (2018)
  37. Survey of hallucination in natural language generation (2023)
  38. Dialogue state tracking with a language model using schema-driven prompting (2021) — link
  39. Competitive Markov decision processes (2012)
  40. Markov decision processes (2013)
  41. Factors affecting text mining based stock prediction: Text feature representations, machine learning models, and news platforms (2022)
  42. Prediction of stock price direction using a hybrid GA-XGBoost algorithm with a three-stage feature engineering process (2022)
  43. A multi-feature stock price prediction model based on multi-feature calculation, LASSO feature selection, and Ca-LSTM network (2022)
  44. Predicting earnings per share using feature-engineered Extreme Gradient Boosting (XGBoost) models (2022)
  45. AutoAlpha: an Efficient Hierarchical Evolutionary Algorithm for Mining Alpha Factors in Quantitative Investment (2020)
  46. Alpha-GPT: Human-AI Interactive Alpha Mining for Quantitative Investment — → 2308.00016
  47. Mining profitable alpha factors via convolution kernel learning (2023)
  48. Enhancing Stock Movement Prediction with Adversarial Training (2019) — link
  49. Mtmd: multi-scale temporal memory learning and efficient debiasing framework for stock trend forecasting (2022) — link
  50. Accurate Multivariate Stock Movement Prediction via Data-Axis Transformer with Multi-Level Contexts (2021) — link
  51. DoubleAdapt: A Meta-learning Approach to Incremental Learning for Stock Trend Forecasting — → 2306.09862
  52. An ensemble of LSTM neural networks for high-frequency stock market classification (2019)
  53. Deep learning with long short-term memory networks for financial market predictions (2018)
  54. A deep learning approach for the Netflix prize (2012) — link
  55. Listwise approach for learning to rank with ranknet (2013)
  56. Prediction of Stock Prices using Machine Learning (Regression, Classification) Algorithms (2020) — link
  57. AutoInt: Automatic feature interaction learning via self-attentive neural networks (2015) — link
  58. Deep learning for information retrieval (2015)
  59. Position-Aware Factorization Machines for Personalized Recommendation (2016)
  60. Bilateral multi-perspective matching for natural language sentences (2017)
  61. Deep ranking using pairwise tensor factorization (2015)
  62. Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (2024)
  63. Fake Alignment: Are LLMs Really Aligned Well? (2023)
  64. Plug in the Safety Chip: Enforcing Constraints for LLM-driven Robot Agents (2023)
  65. PsySafe: A Comprehensive Framework for Psychological-based Attack, Defense, and Evaluation of Multi-agent System Safety (2024)
  66. Universal and Transferable Adversarial Attacks on Aligned Language Models (2023)
  67. Divide-and-Conquer Attack: Harnessing the Power of LLM to Bypass Safety Filters of Text-to-Image Models (2024) — link
  68. ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs (2023) — link
  69. MetaTool Benchmark for Large Language Models: Deciding Whether to Use Tools and Which to Use (2023)
  70. Evil Geniuses: Delving into the Safety of LLM-based Agents (2023)
  71. Watch Out for Your Agents! Investigating Backdoor Threats to LLM-Based Agents (2024)
  72. ToolSword: Unveiling Safety Issues of Large Language Models in Tool Learning Across Three Stages (0)
  73. RoTBench: A Multi-Level Benchmark for Evaluating the Robustness of Large Language Models in Tool Learning (0)
  74. WIPI:s A New Web Threat for LLM-Driven Web Agents (2024) — link
  75. ToRA: A Tool-Integrated Reasoning Agent for Mathematical Problem Solving (2023)
  76. MathCoder: Seamless Code Integration in LLMs for Enhanced Mathematical Reasoning (2023)
  77. Solving Challenging Math Word Problems Using GPT -4 Code Interpreter with Code-based Self-Verification (2024) — link
  78. Measuring Mathematical Problem Solving With the MATH Dataset (2021)

4 of 86 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.